244 Review TRENDS in Cognitive Sciences Vol.5 No.

6 June 2001

Towards a distributed account of

conceptual knowledge
Lorraine K. Tyler and Helen E. Moss

How is conceptual knowledge organized and represented? Are domains (such (primarily visual) and non-sensory functional
as living things) and categories (such as tools, fruit) represented explicitly or properties are represented in functionally and
can domain and category structure emerge out of a distributed system? Taken neuroanatomically separate stores1. Concepts in the
at face value, evidence from brain-damaged patients and neuroimaging studies living domain are more reliant on sensory
suggests that conceptual knowledge is explicitly structured in independent properties whereas artifacts depend on functional
content-based stores. However, recent analyses of the fine-grained details of properties, and therefore, these domains will be
semantic impairments, combined with research using connectionist disproportionately affected by damage to the
modelling, suggest a different picture – one in which concepts are represented sensory and functional systems respectively.
as patterns of activation over multiple semantic properties within a unitary However, the ‘sensory–functional view’ has been
distributed system. Within this context, category-specific deficits emerge as a challenged and two very different alternative
result of differences in the structure and content of concepts rather than from approaches have recently been developed: (1) that
explicit divisions of conceptual knowledge in separate stores. the conceptual system is partitioned into
neuroanatomically distinct content-specific
A fundamental human faculty is our ability to form stores2,3; (2) that category and domain structure are
DOMAINS and CATEGORIES (see Glossary) of knowledge, not explicitly represented but rather are an
to partition objects in the world into meaningful sets emergent property of the structure and content of
such as living things, man-made objects, animals semantic representations4,5.
and tools. During the process of acquiring a new In this article we consider these different
concept, we learn not only its meaning but also the theoretical accounts in relation to behavioural
domain (and category) to which it belongs. One of data, connectionist modelling, lesion data and
the most important distinctions that people seem to neuroimaging studies, and propose that
represent is between the broad domains of living conceptual structure accounts, which focus on the
and non-living things. This is a distinction that is content and structure of concepts, provide a
observed in very young infants, in functional promising theoretical framework for understanding
neuroimaging studies that show selective both the functional and neural organization of
activation in cortical regions associated with conceptual knowledge.
concepts in different domains, and in brain-
damaged patients with deficits restricted to a single Domain/category-specific deficits
domain (e.g. living things, non-living things) or Patients with CATEGORY-SPECIFIC SEMANTIC DEFICITS
category (e.g. animals, fruits, tools) of knowledge. show poorer performance for items in the impaired
Such impairments are typically referred to as domain/category on a range of semantic tasks,
‘category-specific semantic deficits’, even though an including picture naming, word-picture matching,
entire domain of knowledge may be compromised and generation of definitions. Various category-
rather than just a single category. We might assume specific semantic dissociations have been reported
that because domains of knowledge can be including deficits for concrete words compared to
selectively impaired following brain damage and abstract words6,7 and vice versa8, deficits for body
activated in imaging studies they must be explicitly parts9,10, and colour names11. Our major focus here is,
and separately represented in the neural substrate. however, the contrast between the domains of
Whether or not the conceptual system is indeed living/non-living things, because this is the most
organized in this way – as a set of distinct stores of widely studied dissociation. Typically, the impaired
knowledge or in a more distributed system – is a key category is living things12–20, or a specific category
topic in cognitive neuroscience because it relates to within the living-things domain, such as fruit and
the fundamental issue of how cognitive systems in vegetables21,22, or animals2. Only a handful of
general are structured at both the functional and patients have been reported with the reverse pattern
Lorraine K. Tyler*
Helen E. Moss
neural levels. of degraded knowledge specifically for man-made
Dept of Experimental In fact, category-specific deficits have often objects1,16,23–25. Deficits for living things are most
Psychology, University been interpreted as showing that the first-order commonly associated with Herpes Simplex
of Cambridge,
organizing principle in the conceptual system is not Encephalitis (HSE), and have often been linked to
Cambridge, UK CB2 3EB.
*e-mail: lktyler@ domain or category of knowledge but type of bilateral antero-medial and inferior temporal lobe
csl.psychol.cam.ac.uk semantic property – that sensory properties lesions26,27. Deficits for artifacts are claimed to be

http://tics.trends.com 1364-6613/01/$ – see front matter © 2001 Elsevier Science Ltd. All rights reserved. PII: S1364-6613(00)01651-X
 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001 245

not always been observed; some patients with

Glossary
deficits for living things are not impaired on the
Category: refers to individual semantic categories at the superordinate level; for example, supposedly similarly weighted categories of musical
animals, tools, fruit.
Category-specific semantic deficit: following Caramazza29 we use this term to refer to deficits
instruments14, or food15, and there are patients who
that are genuinely semantic in nature – that is, they affect the ability to access conceptual show equal impairments for sensory and functional
knowledge from all modalities of input, and under all output requirements. This review does not properties2,17,31–33. The few patients with artifact
extend to patients whose deficits are restricted to a particular output route (e.g. spoken naming) deficits have rarely been tested in enough detail to
or input route (e.g. visual object recognition) – even though the term category-specific semantic
deficit has often been applied to such patients in the literature. These deficits might be located
determine whether there is a greater deficit for
outside the semantic system. functional properties as predicted by the
Correlation: the degree to which features co-occur in the environment and one feature predicts sensory–functional account, but in two cases where
another (e.g. things that have legs typically also have ears and eyes and can move, breathe and
relevant analyses are reported, the results show
eat; things that have blades are used for cutting and tend to have handles). Within distributed
connectionist systems, properties that co-occur frequently during training support each other
either no difference between sensory and
with mutual activation, so that strongly correlated properties are more resilient to damage than functional properties23, or an effect in the reverse
those which are only weakly correlated. direction32. Finally, the basic premise on which the
Distinctiveness: a measure of how many concepts within a given set contain that property.
sensory–functional hypothesis is predicated has also
A highly distinctive property is specific to only one or a few members of a category. For
example, amongst animals only elephants have a trunk, whereas most animals have legs, and been questioned on the grounds that the contribution
all are alive. of functional information to the concepts of living
Domain: refers to broader groupings, such as living and non-living things. things has been underestimated as the result of an
Semantic properties: components of meaning, as used in the neurospsychological literature.
overly narrow definition of what counts as a
In connectionist modelling, the term feature (or microfeature) is typically used to refer to
semantic properties. For the present review, we draw no distinctions between features functional property. For example, animals have
and properties. important biological functions, such as running,
breathing, eating and so on, that are central to our
conceptual knowledge of this domain (see Tyler and
associated with extensive left-lateralized Moss34 for a discussion) but which tend to be
frontoparietal lesions26. overlooked in analyses of their properties.
Warrington, Shallice and McCarthy1,20 initially
suggested that dissociations between living and Category and domain-specific accounts
non-living things might arise from the differential One response to these problems has been to argue that
contribution of sensory and functional semantic conceptual domain or category, rather than type of
features in distinguishing among concepts within the property (sensory/functional), is, after all, the first
two domains. They claimed that objects (e.g. hammer order organizing principle at both the functional and
and chisel) are more reliably differentiated in terms neuroanatomical level. Caramazza and Shelton2,3
of their function than their sensory properties, have argued this most forcefully, claiming that the
whereas for living things (e.g. cabbage and conceptual system is partitioned into
cauliflower) the reverse holds. On this account the neuroanatomically distinct content-specific domains
conceptual system is fractionated into as a function of their role in the evolution of the human
topographically distinct sub-stores, each of which can brain. Such domains include animals, plant life,
be independently affected by focal brain damage. conspecifics and possibly artifacts. In spite of the
Reports of patients whose deficits cut across prima facie evidence in favour of domain-specific
strict domain boundaries appear to support the representation, caution is needed in accepting this
sensory–functional account. Thus patients who have strong claim for neural segregation. Although the
deficits for living things may also be impaired for lesion data show some broad correlations between
certain non-living categories such as food and musical bilateral antero-medial and inferior temporal lobe
instruments19,20. The explanation is that concepts in damage in patients and living-things deficits26, there
the impaired non-living categories rely on sensory are certainly exceptions2,16. For example, Caramazza
properties in the same way as living things and are and Shelton report patient E.W., who has a selective
therefore affected by damage to the sensory system. deficit for animals in the context of a large area of
In addition, some patients with living-things deficits encephalomalacia within the left posterior frontal and
seem to have a greater impairment for perceptual parietal lobes2. Living-things deficits have also been
than functional properties12,14,19,28, as would be reported in patients with Alzheimer’s dementia, where
predicted on this account. Finally, neuroimaging cortical damage is arguably patchy and widespread4,35.
studies have reported selective activation for animals Also, voxel-based morphometric identification of the
in cortical regions involved in visual analysis, and lesions in patients with HSE show considerable
activation for tools in neural regions which are also asymmetry in the extent to which anterior temporal
activated in action word retrieval30. cortex is compromised in different patients36 with
varying degrees of living-things deficits.
Limitations of a sensory–functional account
However, there are several important problems Evidence from neuroimaging
with the sensory–functional account. The claimed Neuroimaging studies, looking for evidence of neural
associations and dissociations among categories have specialization corresponding to categories or domains

http://tics.trends.com
246 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001

Table 1. Domain-specific activations

Natural kinds minus artifacts Ref.a Artifacts minus natural kinds Ref.
Right inferior parietal lobe (BA 40) 55 Left temporo-occipital junction (BA 37) 55
Bilateral anteromedial temporal lobe 55 Left posterior middle temporal gyrus 55
Left medial occipital gyrus (BA 18) 56 Left posterior middle gyrus (BA 21) 56
Left inferior temporal gyrus (BA 20) 56 Left inferior gyrus (BA 20) 56
Left fourth inferior temporal gyrus 56 Left lateral inferior frontal gyrus (BA 4/6) 30
Left calcarine sulcus 30 Left middle temporal gyrus (BA 21) 30
Left lingual gyrus (BA 18) 38 Right supramarginal gyrus 30
Left fusiform (BA 37) 38 Left anterior cingulate (BA 32) 30
Bilateral anterior temporal (BA 21/38) 37 Left inferior frontal cortex (BA 45) 38
Right posterior middle temporal cortex (BA 37/21) 37 Left posterior temporal cortex (BA 37/19) 37
Left medial extrastriate cortex (BA 18) 37
aTasks in the various studies were as follows. Ref. 55: subjects generated natural kinds or artifacts when presented with a category label, or

words in response to an initial letter. Ref. 56: naming black and white pictures of animals and tools; the baseline task was to judge the
orientation of unfamiliar faces. Ref. 30: silent naming of black and white line drawings of animals or tools, compared with passive viewing of
non-objects. Ref. 38: same/different judgments to pairs of pictures (tools or animals); control conditions were visual texture discrimination
and shape discrimination (meaningless shapes). Ref. 37: silent picture naming and word-picture matching (pictures were either coloured or
black and white); the baseline task was false-font matching.

of knowledge, do not clearly support either the category), we found that conceptual knowledge
domain-specific or sensory–functional accounts. activates a large network, primarily in the LH,
These studies typically test healthy subjects and use a involving the inferior and middle temporal gyri and
variety of different tasks (silent naming, word-picture the temporal pole, and no evidence for regional
matching, category fluency) and materials (pictures, specialization as a function of either category or
silhouettes, words). Although regions of activation domain (see Figs 1 and 2).
specific to living or non-living things have been
identified, these regions are not consistent across The internal structure of categories and concepts
studies (see Table 1). For example, Moore and Price37 Explaining category-specific deficits in terms of
reported activation in bilateral anterior temporal damage to distinct conceptual stores does not in
cortex and right posterior temporal cortex for living itself elucidate the structure and content of the
things, Martin et al.30 found only a small area in the concepts within those stores, and therefore cannot
left calcarine sulcus that was differentially activated provide an adequate explanation for the detailed
for living things, and Perani et al.38 found peak pattern of semantic deficits observed in patients.
activation in the left fusiform and lingual gyrus. Brain damage does not selectively impair a specific
Although these studies found additional processing in type of knowledge or property in an all-or-none
visual cortex for pictures of living things, this might manner. Categories within domains are not always
have been because of their greater visual or equally impaired – for example, musical
structural complexity and so was unrelated to instruments are sometimes categorized with living
category/domain structure per se37. A similar pattern rather than non-living things20. Similarly, the
of inconsistent activations has also been found for vulnerability of concepts within categories varies as
non-living things30,38. a function of many variables, including
The most striking aspect of the neuroimaging data familiarity40, homomorphy, value to perceiver and
is the extent to which living and non-living concepts manipulability41,42. Moreover, some kinds of
activate common regions with only small and properties are more robust to damage than others;
inconsistent differences between domains. These those that are true of many items within a domain
differences might be due to insufficient matching for are generally better preserved than those that are
factors such as frequency, familiarity, imageability more specific2,33, and properties that are densely
and visual complexity. In addition, many studies correlated with each other are better preserved than
report statistical values uncorrected for multiple those that are more weakly inter-related5,43. The
comparisons and thus are liable to false positives. patterns of deficits across many patients suggest
In a series of PET and fMRI studies, where we match that the nature and structure of concepts must play
items across categories and domains on the crucial a crucial role in any theoretical account.
variables of frequency, letter length, and visual An alternative approach attempts to specify how
complexity, we found no differences between the structure of concepts relates to the structure of
categories or domains39. Using a lexical decision task categories and domains and to model apparently
(in which subjects decided whether a letter string selective semantic impairments in a unitary
formed a word) and a semantic categorization task distributed system with no explicit functional or
(in which subjects saw three cue words presented neuroanatomical boundaries according to type of
sequentially and made a speeded decision about concept or property4,5,43–45. This approach assumes
whether a fourth word belonged to the same that concepts are represented within a unitary

http://tics.trends.com
 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001 247

importantly, the contents and structure of the

conceptual representations in the model, specifically
the ways in which relationships between the features
within a concept are captured.
There are two main classes of model: (1) Concepts
are represented by activation over small, intuitively
plausible, experimenter-generated feature sets,
with no claims about the psychological validity of
Fig. 1. Results of PET distributed conceptual system and that the representations. These models assume little
studies using written category-specific deficits emerge as a result of structure in terms of the relations among the
words in semantic
categorization and lexical
differences in the content and structure of concepts features in a concept49,50. (2) Feature structure is
decision tasks. Brain across categories, rather than from explicit divisions emphasized as the major determinant of higher-
areas commonly of conceptual knowledge in independent stores. level structure. These accounts incorporate a set of
activated in the semantic
These claims have recently been explored by theoretically and empirically derived claims about
conditions relative to
their respective baselines considering concepts as patterns of activation over the structure of concepts across the living and
are rendered in an MRI multiple semantic properties within a connectionist non-living domains4,5,43,44,48. The key idea is that
brain image in stereotaxic learning system46–48. concepts in different categories and domains have
space. Red areas are
different internal structures, in terms of the
activated regions
significant at the cluster Connectionist accounts of conceptual structure proportion of features of different kinds
and/or voxel level after Connectionist models of conceptual knowledge (e.g. perceptual versus functional) and in the
correcting for multiple assume that domain and category structure is based CORRELATIONS among features, the relative
comparisons. No voxels
or clusters were
on similarity, captured in the degree to which DISTINCTIVENESS or sharedness of those features over
significantly differentially SEMANTIC PROPERTIES overlap. Thus, lion and canary members of a category and crucially, the many
activated for living things belong to the same domain (living things) because patterns of interaction among these variables. The
or non-living things; that
they share domain-relevant properties (e.g. legs, details of one such model, the conceptual structure
is, there were no effects
of domain (data from move, eyes); however, they fall into different account5,44, are described in Box 1.
Ref. 39). categories (animals versus birds) because some
properties are shared only by concepts within one The effects of brain damage
category (e.g. the wings of a bird) and not by The conceptual structure account claims that
members of the entire domain4,5,43,48,49. Several concepts will be preserved, following brain damage,
models have demonstrated that simple overlap of to the extent that they have numerous
features leads to identifiable clusters in semantic intercorrelated shared properties that support
space corresponding to different categories and/or knowledge of the category or domain as a whole, as
domains49,50. This effect of ‘lumpy semantic space’ well as strong correlations among those more
was predicted by Caramazza et al.45 in their distinctive properties that are necessary for accurate
Organized Unitary Conceptual Hypothesis (OUCH), identification and discrimination among similar
although this was not an implemented model. members of a category. Overall, living things and
Although sharing the same basic principles, artifacts differ along these dimensions. For living
connectionist models differ in terms of specific things, distinctive properties should be vulnerable to
architectures, whether the vector set is based on damage because they are weakly correlated with
hypothetical or real properties, and, most other properties of the concept, whereas shared
properties, being numerous and densely
intercorrelated, should be well-preserved. For
artifacts, the shared–distinctive dissociation should
be less marked, because distinctive properties are
protected by form–function correlations and shared
properties are fewer and less inter-correlated. These
claims predict an interaction between distinctiveness
and domain, with a greater disadvantage for
distinctive compared to shared properties for living
things than artifacts. Different tasks should show
different degrees of dissociation, depending on the
extent to which knowledge of distinctive information
is required. Data supporting these predictions is
reported in Moss et al.33,51
We also predict variation across categories
Fig. 2. Results of an fMRI study using pictures of living and non-living things in a semantic within domains as a function of the structure of the
categorization task showing the neural regions significantly activated in response to (a) artifacts
concepts within each category. For example,
minus baseline, and (b) living-things minus baseline. All activations are masked with the appropriate
contrasts and corrected for multiple comparisons. There were no areas that were significantly according to our property norms, vehicles are like
differentially activated for living things or non-living things. Data from Ref. 57. living things in that they have more numerous and

http://tics.trends.com
248 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001

Box 1. The conceptual structure account

This account combines theoretical correlated with other properties and so are Table I. Global properties of the property
insights and data from neuropsychology, vulnerable to damage. (Note though, that norm set and the model vectors
modelling, developmental and this does not imply that highly distinctive Property Model
experimental psychology with the aim of members of a category will be especially norms vectors
providing an integrated framework in vulnerable. In fact, highly distinctive
which to model normal and disordered members, such as elephants within the Number of concepts 93 96

conceptual systemsa. We claim that living animal category, have many distinctive Highly distinctive 78% 78%
things (most typically animals) have many properties, and this fact increases the features

shared properties (e.g. all mammals probability that at least some of them will Sparsitya 3.7% 4.6%
aDefined as the average proportion of features turned
breathe, have eyes, can see, eat) that be preserved in the face of random
on for each vector.
co-occur frequently and therefore are damage.) By contrast, artifacts have
strongly correlatedb,c. Living things also fewer properties, which tend to be more
have distinctive properties that distinguish distinctive than those of living things. We subsequently tested the validity of our
one category member from another The conceptual structure account also theoretical claims in a property generation
(having stripes versus having spots) incorporates the claim that specific norm study. This confirmed that:
although these tend to be weakly perceptual properties become correlated • Living things have (i) more, (ii) more
with specific functionsd,e. The nature of correlated, and (iii) less distinctive
these form–function correlations properties, than do artifacts.
distinguishes between living things and • Categories within domains differ; for
Semantic output (368)
artifacts: artifacts have distinctive forms example, vehicles are less typical of the
consistently associated with their artifact domain in having more properties
functions (e.g. blade–cut) whereas for overall and a higher ratio of shared to
living things individual variations in form distinctive properties than tools.
tend not to be functionally significant • A higher proportion of correlations
Hidden units (50) (e.g. a lion’s mane)a,b . Even so, living are distinctive for artifacts than for
things (like artifacts) do have living things.
form–function correlations but these We used these property norms to
involve shared properties (e.g. eyes–see; develop a new computational model
legs–move). We refer to these as (see Fig. I), with concepts represented as
Semantic intput (368) biological functionsf,g,. A small-scale vectors instantiating the statistics about
computational model based on these conceptual structure derived from the
claims demonstrated how our property norms (Tables I and II)h. (See Refs i
TRENDS in Cognitive Sciences
assumptions can be captured in a unitary and j for similar approaches.) Vectors were
distributed conceptual system and how presented to the auto-encoder network
Fig. I. Model architecture. The number of units in each
layer is indicated, and arrows show full connectivity ‘lesions’ to such a model can simulate until they all could be recognized (indicated
between layers. category-specific semantic deficitsa,f. by the network reconstructing the input

highly correlated properties than tools, and these it is debatable whether most of these are central
properties are, on average, less distinctive semantic deficits rather than problems with lexical
(H.E. Moss et al., unpublished data). This predicts access or naming10.
that patients who have problems with animals
should also have problems with vehicles in some Organization of semantic space
situations, a pattern we have observed in patients None of the major models of conceptual knowledge
(see Box 2) and which has been reported in at least (domain-specific, sensory–functional and conceptual
one other patient in the literature16 . Our model also structure) can currently account for all of the
predicts that fruit and vegetables will tend to be neuropsychological data. The strength of
particularly vulnerable to damage, since they have connectionist models is that they have the potential
relatively few distinctive properties and these are to account for seemingly complex patterns of
only weakly correlated. This is confirmed by the impaired and preserved features which result in
greater impairment on fruit than animals for category-specific deficits, although they are not yet
several patients in the literature13,17 although there sufficiently well-formulated to account for all of the
are exceptions14. Other selective deficits for data. For example, our current model predicts that
fruit/vegetables versus animals have been reported, artifact deficits will only be seen when damage to
but many of these are restricted to an anomia and so the semantic system is particularly severe5,23,44, and
do not necessarily implicate the semantic system21. thus has difficulty accounting for patients who have
Similarly, highly specific deficits are sometimes greater difficulty with artifacts but in the context of
reported for other categories such as body parts, but a mild impairment16,24.

http://tics.trends.com
 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001 249

Table II. Differences across domains: property norms and model vectors References
a Tyler, L.K. et al. (2000) Conceptual structure
Property norms Model vectors and the structure of concepts: A distributed
Living things Artifacts Living things Artifacts account of category-specific deficits. Brain
Lang. 75, 195–231
Mean no. properties/concept 17.7 11.3 20 14
b Keil, F. (1986) The acquisition of living thing
Mean distinctiveness of properties 0.64 0.73 0.22 0.32 and artifact terms. In Language Learning
No. of shared propertiesb per concept 13.7 7.5 15 6 and Concept Acquisition: Foundational Issues
(Demoupoulous, W. and Marras, A., eds),
bDefined as a property shared by three or more concepts; otherwise the property is distinctive. pp. 133–153, Ablex
c Malt, B.C. and Smith, E. (1984)
Correlated properties in natural categories.
pattern on the output layer); the squared things were more vulnerable to damage J. Verbal Learn. Verbal Behav.
error for each unit was <0.01. The model was than the distinctive properties of artifacts, 23, 250–269
d Madole, K. et al. (1993) Developmental
then ‘lesioned’ to simulate brain damage: whereas the shared properties of living
changes in infants’ attention to function and
connections between layers were randomly things were better preserved than in the case form-function correlations. Cognit. Dev.
removed in 10% increments. As predicted by of artifacts, owing to the greater number of 8, 189–209
the theory, distinctive properties of living shared correlated properties (Fig. II). e De Renzi, E. and Lucchelli, F. (1994) Are
semantic systems separately represented in
the brain? The case of living category
impairment. Cortex 30, 3–25
120 f Durrant-Peatfield, M. et al. (1997) The
distinctiveness of form and function in category
100
structure: a connectionist model. In Proc.
19th Annu. Conf. Cognit. Sci. Soc.
80
(Shafto, M.G. and Langley, P., eds),
% error

60 pp. 193–198, Erlbaum

g Tyler, L.K. and Moss, H.E. (1997) Functional
40 properties of word meaning: studies of normal and
brain-damaged patients. Cognit. Neuropsychol.
20 14, 511–545
h Greer, M. et al. The emergence of semantic
0 categories from distributed featural
0 10 20 30 40 50 60 70 80 90 100 representations. In Proc. 23rd Annu. Conf.
Cognit. Sci. Soc., Erlbaum (in press)
% semantic connections lesioned
i McRae, K. et al. (1997) On the nature and scope
TRENDS in Cognitive Sciences of featural representations of word meaning.
J. Exp. Psychol. Gen. 126, 99–130
Fig. II. The results of ‘lesioning’ the model to simulate brain damage. As predicted, distinctive properties of living j Devlin, J. et al. (1998) Category-specific
things (yellow) were more vulnerable to damage than the distinctive properties of artifacts (blue), whereas the semantic deficits in focal and widespread
shared properties of living things (red) were better preserved than those of artifacts (purple), owing to the greater brain damage: a computational account.
number of shared correlated properties. J. Cogn. Neurosci. 10, 77–94

In spite of their limitations, each of these similarity to other members of the category and
theories provides insights into aspects of the therefore in their distance from the centre of the
neuropsychological data, suggesting that an semantic cluster (see Tranel, Damasio and
account which integrates the properties of each will colleagues for a related account41,42,52,53). Moreover,
be most successful. In our conceptual structure those categories with fewer and less densely inter-
model, we make the assumption that conceptual correlated properties (such as tools or weapons) will
information is randomly distributed without any develop less well-defined regions in semantic space
category/domain organization. This is likely to be an than those for which there is a larger pool of shared,
oversimplification. Given the overlap and structure correlated information (such as animals). In some
in the properties of concepts (Box 1), it is clear that cases, a property might be activated by both an
even a completely distributed connectionist network artifact and a living thing (e.g. ‘used for racing’ is
will develop its own organization such that semantic true of racing cars and greyhounds). Moreover, it is
space will be ‘lumpy’. That is, regions of semantic possible that some categories might deviate from the
space will develop where similar concepts are typical structure for their domain and so exhibit
represented close together by virtue of the fact that non-typical characteristics (e.g. a specific artifact
they share many features which are highly category might have few correlated distinctive
correlated. This will not generate discrete all-or- features). Thus, although the system might organize
none categories of concepts, as overlap of features itself in ways that reflect category and domain
and similarity of structure are probabilistic structure, there will be no discrete independent
variables, with individual concepts varying in their stores corresponding to different categories of

http://tics.trends.com
250 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001

Box 2. Conceptual structure: evidence from patients with category-specific deficits

The results of some recent studies of living things, but the patients do not References
herpes simplex encephalitis patients show this effect (Fig. II). The domain- a Moss, H.E. et al. (1998) ‘Two eyes of a see-through’:
impaired and intact knowledge in a case of selective
with category-specific deficits reveal the specific account does not make any
deficit for living things. Neurocase 4, 291–310
detailed pattern of loss and preservation predictions about the pattern of b Moss, H.E. et al. The emergence of category-
of semantic properties across categories preservation of properties within specific deficit in a distributed semantic system.
and domainsa–d (Figs I–III). These data domains, unless complemented by an In Category-Specificity in Mind and Brain
are consistent with the predictions of additional account of the structure of (Forde, E. and Humphreys, G., eds), Psychology
Press (in press)
the conceptual structure account but concepts. Moreover, it does not readily
c Bunn. E. et al. (1998) Category-specific deficits:
are not readily accommodated by capture the data from the word–picture the role of familiarity re-examined.
other accounts. judgment task (Fig. III), which indicate Neuropsychology 12, 367–379
The sensory–functional account that there is no clear cut-off between an d Tyler, L.K. and Moss, H.E. (1997) Functional
predicts that patients with living-things impaired and a preserved category or properties of word meaning: studies of normal
and brain-damaged patients. Cognit.
deficits will show poorer performance domain, but rather a graded pattern of
Neuropsychol. 14, 511–545
on sensory (perceptual) properties than results that could result from the
functional properties, especially for interaction of the distinctiveness of
properties required to do the task and 100
the conceptual structure of the 90
RC: picture sorting

% correct
100 categories probed. 80
70
90 60
% correct

80 50
40
70 100
30
60 Cross-domain Distant Close
80
50
TRENDS in Cognitive Sciences
60
% correct

40 Fig. III. Word–picture judgment. Patients heard a

100 JBR: picture sorting spoken word (e.g. tiger), which was immediately
90 20 followed by a colour picture on a computer screen, and
% correct

were asked to indicate whether the word and picture

80
0 referred to the same or to different objects. On the
70 ‘different’ trials the picture was either a close category
d e
re iv l co-ordinate (e.g. lion), a distant category co-ordinate
60 a ct na l
Sh in io ua (e.g. mouse) or an object from another domain
50 i st ct ep
t
D Fun rc (e.g. lorry). Close and distant pairs were selected after
Pe extensive pre-testing with control subjects to
determine semantic and visual similarity. Results for a
TRENDS in Cognitive Sciences TRENDS in Cognitive Sciences
group of four HSE patients, including R.C. and J.B.R.,
are plotted. Patients made few errors in the cross-
Fig. I. Picture sorting. Two HSE patients (R.C. and Fig. II. Property verification task. Patients indicated domain condition where they can decide on the basis
J.B.R.) with a well-documented deficit for living-things ‘true’ or ‘false’ to a set of property statements such as of shared properties. They make more errors as the
sorted colour pictures of living and non-living things ‘Tigers have stripes ’ or ‘Spades are used for digging ’. word and picture refer to increasingly similar concepts,
according to two criteria: (1) category level – Statements were equally divided between true and and therefore, increasingly distinctive properties are
animal/bird or tool/vehicle (orange bars); this level of false, perceptual and functional properties, and shared required to differentiate the concepts. There is a clear
sorting depends on shared properties. (2) property versus distinctive properties (from property generation dissociation between living things and tools, especially
level (pale yellow bars); sorting according to specific norms). Results are shown for matched sets of animals in the close condition where the most distinctive
properties, which requires knowledge of distinctive and tools. As predicted by the conceptual structure properties are involved. Note, however, that the pattern
information. At the category level there was no deficit account, patients were more accurate on shared than for vehicles mirrors that of living things rather than
for living things, which according to our theory is distinctive properties, with the poorest performance tools in the close condition. This is consistent with the
because of the better preservation of shared for the distinctive properties of animals. However, no results of our property generation norms, which
properties. The living-things deficit emerged at the patient showed the disadvantage for perceptual indicated that in some respects vehicles are like living
property level, because of the loss of distinctive relative to functional properties of living things, which things, in that they have more numerous and highly
properties of living things relative to artifacts. would be predicted by the sensory–functional account. correlated properties than do tools, and these
Adapted from Ref. a. Adapted from data in Ref. b. properties are on average, less distinctive (see Box 1).

knowledge, but rather graded, overlapping regions the conceptual structure account. This differs from
in semantic space (Fig. 3). Caramazza and Shelton’s domain-specific account2,3
This type of functional organization could be in that these clusters are not independent neural
represented in the neural system as the greater systems that have developed in response to
involvement of specific cortical regions for certain evolutionary pressures. Therefore, even focal brain
categories/domains, such that differential damage will rarely produce all or none deficits for
impairments will arise as a result of local effects, individual categories. If we assume that the
over and above the general patterns of organization of conceptual space at the functional
robustness/vulnerability that have been identified by level is reflected at the neural level, we would expect

http://tics.trends.com
 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001 251

Questions for future research

• Can existing theories of conceptual

knowledge go beyond the representation of
concrete nouns (carrots, cats, corkscrews) to
other word categories (e.g. abstract nouns,
verbs, adjectives)? Are concrete nouns a
special case?
• Connectionist accounts of concept
representation have incorporated a relatively
small number of factors: feature overlap,
sparsity, correlation and distinctiveness. Can
we identify other important determinants of
conceptual structure and model them
TRENDS in Cognitive Sciences successfully in a connectionist framework?
• Can feature-based theories alone account for
Fig. 3. Proposed clustering of correlated features, and the differences conceptual knowledge? Is it necessary to
in structure for concepts in the living and non-living domains, as embed representations of individual concepts
predicted by the conceptual structure model. Each concept is
represented as a pattern of activation over a set of features.
within a framework of theories about the world?
Living-things concepts (green) have many highly intercorrelated If so, what is the relationship between
features (represented by coloured circles) shared by all members of the conceptual and theoretical knowledge?
domain and many intercorrelated properties shared by all members of
• Because most patients with ‘living-things’
a category, such as birds or mammals (represented by light green
versus dark green concepts). Concepts also have some distinctive deficits have antero-medial temporal lobe
features, but these do not tend to be highly correlated with each other. damage, why are these areas not reliably
Artifacts (orange) have fewer, less densely intercorrelated properties at activated in neuroimaging studies of category
either the domain or category level. Therefore categories within the
specificity in normal subjects?
domain (e.g. tools, weapons, vehicles) form less well-defined clusters.
However, distinctive properties tend to occur in small highly
intercorrelated groups; that is, the presence of one property predicts
the presence of another within the concept. In this way, domains and Conclusions
categories form ‘lumps’ within semantic space, but there is no clear
cut-off between them.
Understanding the functional and neural
architecture of the conceptual system remains a
brain damage to produce graded effects; there may be huge, but important, challenge. However,
a certain category or domain of knowledge that is distributed, feature-based accounts of conceptual
Acknowledgements affected more severely than others, but the impact of knowledge provide a promising way forward in
This work was supported the damage will stretch beyond category boundaries terms of explaining the complex patterns of
by an MRC programme to affect other ‘neighbouring’ concepts to various behavioural deficits that arise following brain
grant to L.K. Tyler and
W.D. Marslen-Wilson, and
degrees. Moreover, the pattern of impairment will damage. Moreover, this general approach provides
a Wellcome Trust vary as a function of the processing requirements of a theoretical context within which integrated
Fellowship to H.E. Moss. the specific task at hand (e.g. whether distinctive or accounts of the functional and neural properties of
We thank Rebekah
shared features are emphasized33) and perhaps also conceptual knowledge can be developed. It remains
Anokhina, Mike Greer, and
William Marslen-Wilson by the goal-relevant processing biases associated for future research to determine whether these
for their contributions. with different categories54. promises bear fruit.

References 7 Warrington, E.K. (1981) Concrete word dyslexia. 13 Bunn. E. et al. (1998) Category-specific deficits:
1 Warrington, E.K. and McCarthy, R. (1987) Br. J. Psychol. 72, 175–196 the role of familiarity re-examined.
Categories of knowledge: further fractionations and 8 Tyler, L.K. et al. (1995) Abstract word deficits in Neuropsychology 12, 3, 367–379
an attempted integration. Brain 110, 1273–1296 aphasia: evidence from semantic priming. 14 De Renzi, E. and Lucchelli, F. (1994) Are
2 Caramazza, A. and Shelton, J.R. (1998) Neuropsychology 9, 354–363 semantic systems separately represented in the
Domain-specific knowledge systems in the brain: 9 Dennis, M. (1976) Dissociated naming and brain? The case of living category impairment.
the animate–inanimate distinction. J. Cogn. locating of body parts after left anterior temporal Cortex 30, 3–25
Neurosci. 10, 1–35 lobe resection: an experimental case study. Brain 15 Hart, J. and Gordon, B. (1992) Neural subsystems
3 Shelton, J.R and Caramazza, A. (1999) Deficits in Lang. 3, 147–163 for knowledge. Nature 359, 60–64
lexical and semantic processing: implications for 10 Suzuki, K. et al. (1997) Category-specific 16 Hillis, A.E. and Caramazza, A. (1991)
models of normal language. Psychonomic Bull. comprehension deficit restricted to body parts. Category-specific naming and comprehension
Rev. 6, 5–27 Neurocase 3, 193–200 impairment: a double dissociation. Brain Lang.
4 Gonnerman, L. et al. (1997) Double dissociation 11 DeVreese, L. (1988) Category-specific versus 114, 2081–2094
of semantic categories in Alzheimer’s Disease. modality-specific aphasia for colours: a review of 17 Laiacona, M. et al. (1997) Semantic category
Brain Lang. 57, 254–279 the pioneer case studies. Int. J. Neurosci. dissociations: a longitudinal study of two cases.
5 Tyler, L.K. et al. (2000) Conceptual structure and 43, 195–206 Cortex 33, 441–461
the structure of concepts. Brain Lang. 75, 195–231 12 Basso, A. et al. (1988) Progressive language 18 Moss, H.E. et al. (1997) When leopards lose their
6 Breedin, S.D. et al. (1994) Reversal of the impairment without dementia: a case with spots: knowledge of visual properties in
concreteness effect in a patient with semantic isolated category specific semantic impairment. category-specific deficits for living things.
dementia. Cognit. Neuropsychol. 11, 617–660 J. Neurol. Neurosurg. Psychiatry 51, 1201–1207 Cognit. Neuropsychol. 14, 901–950

http://tics.trends.com
252 Review TRENDS in Cognitive Sciences Vol.5 No.6 June 2001

19 Sartori, G. and Job, R. (1988) The oyster with knowledge? Evidence from a category-specific 45 Caramazza A. et al. (1990) The multiple
four legs: a neuropsychological study on the double dissociation. Neurocase 4, 311–338 semantics hypothesis: multiple confusions?
interaction of visual and semantic information. 33 Moss, H.E. et al. (1998) ‘Two eyes of a see-through’: Cognit. Neuropsychol. 7, 161–189
Cognit. Neuropsychol. 5, 105–132 impaired and intact knowledge in a case of selective 46 Hinton, G.E and Shallice, T. (1991) Lesioning an
20 Warrington, E.K. and Shallice, T. (1984) deficit for living things. Neurocase 4, 291–310 attractor network: investigations of acquired
Category specific semantic impairments. Brain 34 Tyler, L.K. and Moss, H.E. (1997) Functional dyslexia. Psychol. Rev. 98, 74–95
107, 829–854 properties of word meaning: studies of normal and 47 Masson, M.E.J. (1995) A distributed memory
21 Hart, J. et al. (1985) Category-specific naming deficit brain-damaged patients. Cognit. Neuropsychol. model of semantic priming. J. Exp. Psychol.
following cerebral infarct. Nature 316, 439–440 14, 511–545 Learn. 21, 3–23
22 Sheridan, J. and Humphreys, J.W. (1993) A verbal 35 Silveri, M.C. et al. (1991) Dissociation between 48 McRae, K. et al. (1997) On the nature and scope
semantic category-specific recognition deficit. knowledge of living and nonliving things in of featural representations of word meaning.
Cognit. Neuropsychol. 10, 143–184 dementia of the Alzheimer type. Neurology J. Exp. Psychol. Gen. 126, 99–130
23 Moss, H.E. and Tyler, L.K. (2000) A progressive 41, 545–546 49 Small, S. et al. (1995) Distributed representations
category-specific deficit for non-living things. 36 Gitelman, D. et al. Voxel-based morphometry of of semantic knowledge in the brain. Brain
Neuropsychologia 38, 60–82 Herpes Simplex Encephalitis. NeuroImage 118, 441–453
24 Sacchett, C. and Humphreys, G. (1992) Calling (in press) 50 Zorzi, M. et al. (1999) Category-specific deficits in
a squirrel a squirrel but a canoe a wigwam: a 37 Moore, C.J. and Price, C.J. (1999) A functional a self-organizing model of the lexical-semantic
category-specific deficit for artifactual objects neuroimaging study of the variables that generate system. In Connectionist Models in Cognitive
and body parts. Cognit. Neuropsychol. 73–86 category-specific object processing differences. Neuroscience (Heincke, D. et al., eds),
25 Warrington, E.K. and McCarthy, R. (1983) Category Brain 122, 943–962 pp. 137–148, Springer-Verlag
specific access dysphasia. Brain 106, 859–878 38 Perani, D. et al. (1995) Different neural systems 51 Moss, H.E. et al. The emergence of category
26 Gainotti, G. et al. (1995) Neuroanatomical for the recognition of animals and man-made specific deficits in a distributed semantic system.
correlates of category-specific impairments: tools. NeuroReport 6, 1637–1641 In Category-Specificity in Brain and Mind
a critical survey. Memory 3/4, 247–264 39 Devlin, J.T. et al. (2000) Susceptibility induced (Forde, E. and Humphreys, G.W., eds), Psychology
27 Pietrini, V. et al. (1988) Recovery from herpes loss of signal: comparing PET and fMRI on a Press (in press)
simplex encephalitis: Selective impairment of semantic task. NeuroImage 11, 589–600 52 Damasio, A.R. et al. (1990) Category-related
specific semantic categories with 40 Funnell, E. and Sheridan, J. (1992) Categories recognition deficits as a clue to the neural
neuroradiological correlation. J. Neurol. of knowledge: unfamiliar aspects of living and substrates of knowledge. Trends Neurosci.
Neurosurg. Psychiatry 51, 1284–1293 non-living things. Cognit. Neuropsychol. 13, 95–98
28 Silveri, M.C. and Gainotti, G. (1988) 9, 135–153 53 Damasio, A.R. and Damsio, H. (1993)
Interaction between vision and language in 41 Tranel, D. et al. (1997) Explaining category Cortical systems underlying knowledge
category-specific semantic impairment. Cognit. related effects in the retrieval of conceptual and retrieval: evidence from human lesion
Neuropsychol, 5, 677–709 lexical knowledge for concrete entities: studies. In Exploring Brain Functions:
29 Caramazza, A. (2000) The organization of operationalization and analysis of factors. Models in Neuroscience (Poggio, T.A. and
conceptual knowledge in the brain. In The Neuropsychologia 35, 1329–1339 Glaser, G.A., eds), pp. 233–248,
New Cognitive Neurosciences (2nd edn) 42 Tranel, D. et al. (1997) A neural basis for the John Wiley & Sons
(Gazzaniga, M.S., ed.) pp. 1037–1046, retrieval of conceptual knowledge. 54 Gauthier, I. (2000) What constrains the
MIT Press Neuropsychologia 35, 1319–1327 organization of the ventral temporal cortex.
30 Martin, A. et al. (1996) Neural correlates of 43 Devlin, J.T. et al. (1998) Category-specific Trends Cognit. Sci. 4, 1–2
category-specific knowledge. Nature semantic deficits in focal and widespread brain 55 Mummery, C.J. et al. (1996) Generating ‘tiger’ as
379, 649–652 damage: a computational account. J. Cogn. an animal name or a word beginning with T:
31 Funnell, E. and de Mornay Davies, P. (1996) JBR: Neurosci. 10, 77–94 differences in brain activation. Proc. R. Soc.
a reassessment of concept familiarity and a 44 Durrant-Peatfield, M. et al. (1997) The London Ser. B 263, 989–995
category-specific disorder for living things. distinctiveness of form and function in category 56 Damasio, H. et al. (1996) A neural basis for lexical
Neurocase 2, 461–474 structure: a connectionist model. In Proceedings retrieval. Nature 380, 499–505
32 Lambon Ralph, M.A. et al. (1998) Are living and of the Nineteenth Annual Conference of the 57 Tyler, L.K. and Tavares, P. Is there neural
non-living category-specific deficits causally Cognitive Science Society (Shafto, M.G. and specialization for pictures of living and non-living
linked to impaired perceptual or associative Langley, P., eds), pp. 193–198, Erlbaum things? NeuroImage (in press)

Book reviews
Books and other types of publication – including software, CD-ROMs, films, art exhibitions, etc. – can be
considered for review in Trends in Cognitive Sciences. Please send any ideas or proposals for specific
items to the Editor.
Publishers: send copies of books or software you would like to be reviewed to the Editor at the address
below.
Readers: would you like to review a particular book or other publication? Contact the Editor via e-mail with
your proposal.
Trends in Cognitive Sciences
Elsevier Science London, 84 Theobald’s Road, London WC1X 8RR.
e-mail: [email protected]

http://tics.trends.com